A novel physics-constrained deep learning framework for the inverse design of assembly contact interfaces

Assembly contact interface characteristics critically influence the performance of precision mechanical systems. Traditional design methods relying on iterative finite element analysis are computationally expensive, while existing deep learning approaches often neglect physical constraints and the complex effects of assembly processes. To address these limitations, this paper proposes a physics-constrained deep learning framework for the inverse design of assembly interfaces. Specifically, we introduce a novel network architecture which integrates multi-source inputs including target contact pressure, assembly parameters, and service conditions. To enforce physical consistency, a differentiable loss function incorporating the impenetrability condition is developed. Furthermore, an optimized learning rate scheduling strategy is implemented to enhance model convergence. Comprehensive ablation and comparative experiments demonstrate that our method outperforms conventional approaches in both accuracy and physical plausibility. When applied to an aero-engine flange structure, the framework enables rapid inverse design of interface morphology, reducing maximum contact pressure by 15.67% and increasing the effective contact area by 45.23% compared to traditional designs. This work provides a robust solution for assembly interface design and advances the application of physics-constrained deep learning in complex engineering systems.